Fix signature of torch.sparse_coo_tensor() #152681
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/152681
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit a6f73ec with merge base 3443627 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@pytorchbot label "release notes: sparse" |
There was a problem hiding this comment.
Thanks for the PR, @ILCSFNO!
This PR eliminates the TypeError reported in #145371 and is a right step forward fixing the issue but we are not quite there yet.
To be specific, the behavior of sparse_coo_tensor with and without the size argument should be identical. With this PR, there exists a couple of discrepancies. For instance, consider the following sample data
# uncoalesced indices
i = torch.tensor([[0, 1, 0], [1, 2, 0]])
# coalesced indices
i2 = torch.tensor([[0, 0, 1], [0, 1, 2]])
# values
v = torch.tensor([3.0, 4.0, 5.0])
s = (2, 3)then:
Inconsistency 1
# succeeds, UNEXPECTED:
result1 = torch.sparse_coo_tensor(i, v, check_invariants=True, is_coalesced=True)
# fails, EXPECTED:
result2 = torch.sparse_coo_tensor(i, v, s, check_invariants=True, is_coalesced=True)
# -> RuntimeError: cannot set is_coalesced to true if indices correspond to uncoalesced COO tensorInconsistency 2
result1 = torch.sparse_coo_tensor(i, v, check_invariants=False, is_coalesced=True)
result2 = torch.sparse_coo_tensor(i, v, s, check_invariants=False, is_coalesced=True)
assert result1.is_coalesced() == result2.is_coalesced() # FAILSAlso a related bug:
result1 = torch.sparse_coo_tensor(i2, v, check_invariants=True, is_coalesced=True)
result2 = torch.sparse_coo_tensor(i2, v, s, check_invariants=True, is_coalesced=True)
assert result1.is_coalesced() == result2.is_coalesced() # FAILS|
@pearu Fixed ready for your check again! Thanks! |
Fixes #145371
@pearu Searched all and find these codes, wondering whether is the root cause of the issue, could you have a review? Thanks a lot!